DeepSeek swapped its Flash model overnight. I pulled the new config and did the cache math.

On October 17, DeepSeek replaced the model behind its deepseek-flash API endpoint. The new checkpoint, DeepSeek-V4.1-Flash-1017, is now available on Hugging Face under MIT license. The model shows improvements in several benchmarks, such as Terminal-Bench 4.0 and ProgramBench, with an increase of 11.1 and 8.8 points respectively.

However, the KV cache compression paper suggests a cache cost of 890 bytes per token, but the actual config reveals a more complex situation. The model uses a sliding window of 128 tokens, storing only four source layers for long-range KV, with compression ratios on top. When analyzing the cache cost, it becomes clear that the figure of 890 bytes per token is not a constant for this architecture, and it varies depending on the context length.

This can range from 688 bytes per token at 8,192 tokens, to 43 bytes per token at 131,072 tokens, and down to 5.4 bytes per token at 1,048,576 tokens. DeepSeek's API pricing demonstrates the cost difference between uncached and cached input tokens, with cached input being 50 times cheaper. This knowledge sheds light on the model's efficiency and the vendor's strategy of offering low-cost cache reads.